[RFC] Host1x/TegraDRM UAPI

[Mikko Perttunen]

On 6/25/20 3:47 AM, Dmitry Osipenko wrote:
> 23.06.2020 15:09, Mikko Perttunen пишет:
>> #define DRM_TEGRA_SUBMIT_CREATE_POST_SYNC_FILE      (1<<0)
>> #define DRM_TEGRA_SUBMIT_CREATE_POST_SYNCOBJ        (1<<1)
> 
> The sync object shouldn't be created by the kernel driver and we
> shouldn't need the sync-file FD.
> 
> Please consider this example of how syncobj may be used:
> 
>    1. Syncobj is created by userspace.
> 
>    2. Syncobj's handle (post_fence) is passed with the job to the kernel
> driver.
> 
>    3. Userspace waits on syncobj for the job's submission.
> 
>    4. Kernel driver attaches job-completion dma-fence(s) to the syncobj
> after starting to execute the job.
> 
>    5. Userspace waits on syncobj for the job's completion.
> 
> Syncobj is a superset of the sync-file fence:
> 
>    a) Syncobj doesn't have a backing file descriptor when it's created.
> For example, why would you need an FD if you're not going to share fence
> with other processes. It's possible to get FD out of syncobj later on,
> please see [1][2].
> 
>    b) Syncobj is designed to be used with a GPU jobs. For example,
> userspace passes job to the GPU driver's scheduler and then waits until
> job has been started to execute on hardware, this is already supported
> by syncobj.
> 
>    c) It is possible to attach sync-file's fence to a syncobj [1][2][3]
> and now:
> 
>    - Userspace may attach sync-file's fence to a syncobj.
> 
>    - Userspace may use this syncobj for the job's pre-fence.
> 
>    - Kernel driver will take out the pre-fence from the syncobj during of
> the job's submission and reset the syncobj's fence to NULL.
> 
>    - Job's scheduler will wait on this pre-fence before starting to
> execute job.
> 
>    - Once the job is started to execute, the job's scheduler will attach
> the job's completion-fence to the syncobj. Now syncobj has a post-fence!
> 
>    d) It is possible to get sync-file FD out of syncobj [1][2][4].
> 
> [1]
> https://elixir.bootlin.com/linux/v5.7.2/source/include/uapi/drm/drm.h#L730
> [2]
> https://elixir.bootlin.com/linux/v5.7.2/source/include/uapi/drm/drm.h#L933
> [3]
> https://elixir.bootlin.com/linux/v5.7.2/source/drivers/gpu/drm/drm_syncobj.c#L653
> [4]
> https://elixir.bootlin.com/linux/v5.7.2/source/drivers/gpu/drm/drm_syncobj.c#L674
> 
> ===
> 
> So, sync object can carry both pre-fence and post-fence, and they could
> be sync-file FDs!
> 
> The corollary is that we can get away by using a single syncobj handle
> for the job's submission IOCTL.
> 

Ah, clearly I hadn't understood syncobjs properly :) Last time I spent 
significant time with this they didn't exist yet.. I'll check this out.

Mikko